Pass-through where it counts. A platform on top.
OpenAI, Anthropic, and Google ship model APIs. Ingram Cloud runs agents over all three and adds the layer they leave to you. Because we speak the OpenAI wire format, most model features come along for free — and the rows the providers can’t fill are the ones that turn a model into a product.
Reflects each provider’s first-party API as of mid-2026. The model APIs move fast; the Ingram Cloud column is checked against the live surface.
| Feature | OpenAI | Anthropic | Gemini | Ingram Cloud |
|---|---|---|---|---|
Core inference The chat surface itself. We speak the OpenAI wire format and route it to whichever model family the agent runs on. | ||||
OpenAI Chat Completions format point the openai SDK or AI SDK straight at us | ~ | ~ | ||
Responses API surface | ||||
Streaming (SSE) | ||||
Sampling controls temperature, top_p, max_tokens, stop, seed | Soon | |||
Token & cost usage reporting Ingram Cloud meters per smith, per customer | ||||
Tools & function calling Client-side function loops pass straight through; server-side tools run on our side over MCP, with approval gating the providers don't offer. | ||||
Function / tool calling | ||||
Parallel tool calls | ||||
Forced tool choice auto / none / required / a named tool | ||||
Remote MCP tools Ingram Cloud connects once, server-side, default-deny | ||||
Hosted web search | ||||
Code interpreter / sandbox | Soon | |||
Human approval gates pause the run before anything irreversible | ~ | |||
Structured output & inputs Schema-constrained output and multimodal inputs, passed through to the model. | ||||
JSON-schema structured output | ~ | |||
Image input (vision) | ||||
PDF / document input inline bytes stored for audit, fetchable by id | ||||
Generative media Beyond text. These are provider model capabilities — Ingram Cloud focuses on the assistant loop, not media generation. | ||||
Audio input | ||||
Speech output / TTS | ||||
Realtime voice speech-to-speech | ||||
Image generation | ||||
Video understanding | ||||
Retrieval, batch & safety The surrounding model-API surface — some passed through, some on our roadmap. | ||||
Embeddings API we embed internally for recall; not yet a public endpoint | Soon | |||
Reranking | Soon | |||
Content moderation | ~ | ~ | ||
Batch API | Soon | |||
Prompt caching provider-default; we track and bill cache reads/writes | ~ | |||
Fine-tuning | ~ | |||
File upload API inline files work today; a first-class upload endpoint is next | Soon | |||
Agent platform & governance The layer the model APIs don't attempt — and the reason to build on Ingram Cloud over any single provider. | ||||
All three model families, one API OpenAI, Claude, and Gemini behind a single token | ||||
Bring-your-own or hosted keys | ||||
Versioned agent designs & rollout | ~ | ~ | ||
A private instance per end-user the smith — isolated memory, tools, identity | ||||
Server-side per-user memory & recall | ||||
Durable runs (pause / resume / replay) | ~ | |||
Scheduled / cron runs | ||||
Channel delivery Slack, WhatsApp, Telegram, email | ||||
Event feed & signed webhooks | ~ | |||
Per-end-user budgets & metering | ||||
Full run trace / audit log | ~ | |||
EU data residency | ~ | ~ | ||
Everything the model APIs do — plus the part they don't.
Keep the SDK you already use, reach all three model families through one token, and get the per-user memory, approvals, channels, and audit trail the providers leave to you.